A fault diagnosis method for motor vibration signals incorporating Swin transformer with locally sensitive hash attention

Abstract Identification of motor vibration signals is one of the important tasks in motor fault diagnosis and predictive maintenance, and wavelet time–frequency diagram is a commonly used signal analysis method to extract the frequency and time characteristics of signals. In this paper, a method based on local sensitive hashing (LSH)-Swin transformer network is proposed for identifying the wavelet time–frequency diagrams of motor vibration signals to analyze the fault types. The traditional Swin transformer model converges slowly due to the smoothing of the attention distribution when dealing with data with sparse features, while the method proposed in this paper reduces the smoothing of the computed attention and enables the network to learn the key features better by introducing locally-sensitive hash attention in the network model, dividing the sequences in the input attention into multiple hash buckets, calculating the attention weights of only some of the vectors with a high degree of hash similarity, and by sampling discrete samples with the use of the Gumbel Softmax. The experimental results show that the method proposed in this paper has better recognition accuracy and higher computational efficiency compared with the traditional network when processing wavelet time–frequency maps of motor vibration signals, and its validation accuracy reaches 99.7%, the number of parameters also has a decrease of about 13%, and the training network to reach converged epochs is also faster. The method in this paper can provide an effective solution for the analysis and processing of motor vibration signals, and has certain application value in practical engineering.